The zero-width myth won't die
Within days of Anthropic announcing watermarking, a dozen tools appeared promising to remove it. Nearly all of them do the same thing: strip zero-width Unicode characters.
Claude does not use zero-width characters. The tools are cleaning something real, but it is not the watermark.
Where the idea came from
The zero-width story predates this announcement by years. It circulated as folklore about AI detection — the claim that chatbots quietly inserted invisible characters as trackers.
It was easy to demonstrate, easy to believe, and easy to build a product around. When watermarking became real news, the existing story was simply attached to it.
Why it cannot be true
A watermark that any user can delete with a one-line regular expression provides no provenance guarantee at all.
Anthropic built this to satisfy a transparency obligation. A trivially removable mark would not satisfy anything, and the company's own description — a watermark woven into the text itself — rules it out.
The part that is genuinely useful
Character scrubbing is still worth doing. Pasted text picks up stray formatting characters that break search, diffing and sometimes code.
The problem is not the tool. It is the label on the tool, and the false confidence it produces in people who then publish believing they have handled something they have not.